Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
1. A processor implemented method ( 300 ) comprising: receiving a plurality of images, at one or more time intervals, pertaining to a context under consideration, the received images being correlated and associated with at least one of a spatial and temporal information ( 302 ); identifying and transmitting, at the one or more time intervals, at least a subset of the received images based on the spatial or temporal information and an adaptive threshold ( 304 ); extracting features from the received images, by forward passing the received images through a neural network model pre-trained on a dataset of a plurality of images pertaining to varied contexts ( 306 ), wherein the step of extracting features from the received images is preceded by a step of pre-processing comprising at least one of (i) a first level of pre-processing, to enhance quality of the received images, by performing one or more of normalization, Principal Components Analysis (PCA) whitening, brightness correction, standardization and segmentation; and (ii) a second level of pre-processing, to adapt the received images for forward passing to the neural network model, by performing one or more of rotation, cropping, shifting, scaling and zooming; performing a first level of inferencing ( 308 ), by clustering the received images into one or more classes using the extracted features, the step of clustering comprising: determining an optimal number of the one or more classes using a Silhouette coefficient ( 308 a ); identifying the one or more classes based on similarity detected between the received images ( 308 b ) by performing at least one of: computing a first distance measure represented by a spherical distance of each of the received images with every other image in the received images ( 308 b - 1 ); computing a likelihood of each image in the received images to belong to a class using a class distribution based on a Maximum A Posteriori probability (MAP) ( 308 b - 2 ); and computing a second distance measure represented by a spherical distance between each of the received images and a centroid of each of the one or more classes ( 308 b - 3 ); wherein the number of the one or more classes equals the determined optimal number; and validating the quality of the one or more classes using one or more of an Normalized Mutual Information (NMI) score, a Rand Index and a purity measure ( 308 c ); and associating the one or more classes with a tag based on the context under consideration ( 310 ).
This invention relates to a processor-implemented method for analyzing and clustering images based on spatial and temporal information. The method addresses the challenge of efficiently processing large datasets of images to extract meaningful patterns and context-specific insights. The system receives multiple images at different time intervals, each associated with spatial or temporal metadata. An adaptive threshold is used to select a subset of images for further analysis, optimizing computational efficiency. Before feature extraction, the images undergo two levels of preprocessing: the first enhances image quality through normalization, PCA whitening, brightness correction, standardization, and segmentation, while the second adapts the images for neural network input via rotation, cropping, shifting, scaling, and zooming. A pre-trained neural network model processes the preprocessed images to extract features. These features are then used in a clustering step, where the optimal number of classes is determined using the Silhouette coefficient. Clustering is performed by computing spherical distances between images, evaluating class likelihoods via Maximum A Posteriori probability, and measuring distances to class centroids. The quality of the clustering is validated using metrics like Normalized Mutual Information, Rand Index, and purity. Finally, the identified classes are tagged based on the context under consideration, enabling contextual analysis and interpretation of the image data.
2. The processor implemented method of claim 1 , wherein the step of receiving comprises obtaining at least some metadata associated with the received images; and updating the metadata associated with the received images based on the associated tags.
The invention relates to image processing systems that enhance metadata management for digital images. The problem addressed is the lack of efficient methods to automatically update and maintain metadata associated with images, particularly when tags or labels are applied to those images. Current systems often require manual intervention or lack integration between tagging and metadata updates, leading to inconsistencies and inefficiencies. The method involves receiving digital images along with their associated metadata, which may include information such as timestamps, author details, or descriptive tags. The system then processes these images to extract or apply additional tags, which are labels or keywords that describe the content or context of the images. Once the tags are identified or assigned, the system updates the metadata of the images to reflect these tags, ensuring that the metadata remains accurate and up-to-date. This automated process eliminates the need for manual metadata updates and improves the consistency and reliability of image metadata in databases or storage systems. The method is particularly useful in applications such as digital asset management, content moderation, and automated image cataloging.
3. The processor implemented method of claim 2 , wherein the step of extracting features from the received images comprises one or more of appending the features extracted by the neural network model, with additional features including morphological features and color related features to derive a master feature set; and compressing the master feature set using dimensionality reduction methods.
This invention relates to image processing and feature extraction techniques, particularly for enhancing the accuracy and efficiency of image analysis using neural networks. The method addresses the challenge of improving feature representation in images by combining multiple feature types and optimizing their dimensionality. The process involves receiving one or more images and extracting features using a neural network model. To enhance the feature set, additional features are appended, including morphological features (e.g., shape, texture) and color-related features (e.g., histograms, color spaces). These supplementary features are combined with the neural network-derived features to form a comprehensive master feature set. The master feature set is then compressed using dimensionality reduction techniques, such as principal component analysis (PCA) or autoencoders, to retain only the most relevant information while reducing computational overhead. This approach improves image analysis by leveraging diverse feature types and optimizing their representation, making it suitable for applications like object recognition, medical imaging, and autonomous systems where robust and efficient feature extraction is critical.
4. The processor implemented method of claim 3 , further comprising the step of performing a second level of inferencing by classifying the one or more classes into one or more sub-classes ( 312 ), the step of classifying comprising: obtaining a plurality of pre-trained tagger models associated with one or more sub-classes corresponding to one or more classes pertaining to the context under consideration, wherein the pre-trained tagger models are trained by the master feature set ( 312 a ); classifying the one or more classes from the first level of inferencing into the one or more sub-classes based on the plurality of pre-trained tagger models ( 312 b ); computing a confidence level for the classified one or more sub-classes ( 312 c ); re-training the plurality of pre-trained tagger models with the received images associated with the confidence level below a pre-defined threshold to obtain a plurality of classification models ( 312 d ); and creating a knowledge ontology of the classifications models based on the one or more classes, the one or more sub-classes therein and the inter-relationships thereof ( 312 e ).
This invention relates to a hierarchical classification system for image analysis, specifically addressing the challenge of accurately categorizing images into fine-grained sub-classes beyond broad initial classifications. The method involves a two-level inferencing process. In the first level, images are classified into broad classes using a master feature set. The second level further refines these classes by sub-classification. This is achieved by obtaining pre-trained tagger models, each associated with specific sub-classes relevant to the context of the analysis. These models are trained using the master feature set. The system then classifies the initial broad classes into sub-classes based on the tagger models and computes confidence levels for these sub-classifications. If the confidence level falls below a predefined threshold, the tagger models are re-trained using the corresponding images to improve accuracy. Finally, the system creates a knowledge ontology that maps the relationships between the broad classes, sub-classes, and their interconnections. This approach enhances the precision of image classification by leveraging hierarchical refinement and adaptive model training.
5. The processor implemented method of claim 4 , wherein the step of re-training the plurality of pre-trained tagger models is preceded by evaluating the received images associated with the confidence level below the pre-defined threshold with the associated metadata.
This invention relates to improving the accuracy of image tagging systems by re-training pre-trained tagger models. The problem addressed is the occurrence of low-confidence image tags, which can degrade the performance of automated tagging systems. The solution involves evaluating images with low-confidence tags (below a predefined threshold) along with their associated metadata to identify errors or inconsistencies. Based on this evaluation, the pre-trained tagger models are re-trained to improve their accuracy. The re-training process may involve adjusting model parameters, updating training datasets, or refining tagging algorithms to better handle ambiguous or low-confidence cases. The method ensures that the tagging system continuously improves by learning from its mistakes, particularly focusing on images where the initial tagging was uncertain. This approach enhances the reliability of automated image tagging in applications such as content moderation, search engines, or digital asset management. The system may also incorporate feedback loops where human reviewers validate the re-trained models' performance, further refining the tagging process. The overall goal is to minimize low-confidence tags and improve the consistency and accuracy of automated image classification.
6. The processor implemented method of claim 5 further comprising storing the received images and the metadata associated thereof; the one or more classes; the one or more sub-classes, the plurality of classification models and the knowledge ontology ( 314 ).
This invention relates to a processor-implemented method for managing and classifying images and associated metadata using machine learning models and a knowledge ontology. The method addresses the challenge of efficiently organizing and retrieving large volumes of image data by leveraging structured classification hierarchies and domain-specific knowledge representations. The method involves receiving images and their associated metadata, which may include descriptive tags, timestamps, or other contextual information. These images are then classified into one or more predefined classes and further refined into sub-classes based on their content. The classification process utilizes a plurality of classification models, each trained to recognize specific features or patterns within the images. These models may employ deep learning techniques, such as convolutional neural networks, to analyze visual data and assign appropriate classifications. A knowledge ontology is used to structure the relationships between classes, sub-classes, and metadata, enabling semantic reasoning and improved data retrieval. The ontology defines hierarchical relationships, attributes, and constraints that enhance the interpretability and usability of the classified data. The method also includes storing the received images, their metadata, the classification models, and the knowledge ontology in a structured database for future reference and analysis. This approach improves the efficiency of image classification and retrieval by integrating machine learning with structured knowledge representation, making it particularly useful in applications such as medical imaging, surveillance, or content-based image retrieval systems.
7. The processor implemented method of claim 6 , wherein the step of identifying and transmitting at least a subset of the received images comprises performing one or more of: determining whether the received images are valid based on an entropy value associated thereof; and comparing the received images with a pre-determined number of previous images to determine whether the received images fit into the one or more sub-classes associated thereof, wherein the pre-determined number of previous images is the adaptive threshold based on the context under consideration.
This invention relates to image processing methods for validating and classifying received images in a processor-implemented system. The method addresses the problem of efficiently identifying and transmitting relevant subsets of images while filtering out invalid or redundant data. The process involves determining the validity of received images by analyzing their entropy values, ensuring only high-quality or meaningful images are processed further. Additionally, the method compares received images with a predetermined number of previous images to assess whether they fit into predefined sub-classes. The number of previous images used for comparison is dynamically adjusted based on the context, allowing for adaptive thresholding that improves classification accuracy. This adaptive approach ensures that the system can handle varying conditions and image types effectively, optimizing both storage and transmission efficiency. The method is particularly useful in applications requiring real-time image analysis, such as surveillance, medical imaging, or autonomous systems, where minimizing unnecessary data processing is critical. By combining entropy-based validation with context-aware adaptive thresholding, the invention enhances the reliability and efficiency of image classification systems.
8. The processor implemented method of claim 1 , wherein the step of performing the first level of inferencing is performed within one or more of the one or more classes.
The invention relates to a processor-implemented method for optimizing machine learning inferencing within a classification system. The method addresses the computational inefficiency of traditional inferencing approaches, which often require processing all possible classes in a dataset, leading to unnecessary resource consumption and slower performance. The method involves a hierarchical inferencing process where a first level of inferencing is performed within one or more specific classes. This means that instead of evaluating all classes indiscriminately, the system narrows down the inferencing scope to only the relevant classes, reducing the computational load. The first level of inferencing may include preliminary filtering or scoring to identify the most likely candidates before proceeding to deeper analysis. This selective processing improves efficiency by avoiding unnecessary computations on irrelevant classes. The method may also include additional inferencing levels, where each subsequent level further refines the results based on the outputs of the previous level. This multi-stage approach ensures that only the most relevant classes are fully processed, optimizing both speed and accuracy. The system dynamically adjusts the inferencing scope based on the input data, ensuring adaptability to different scenarios. By performing the first level of inferencing within specific classes rather than across all classes, the method significantly reduces the computational overhead while maintaining or improving inferencing accuracy. This approach is particularly useful in large-scale classification tasks where processing all classes would be impractical.
9. A system ( 200 ) comprising: one or more internal data storage devices operatively coupled to one or more hardware processors for storing instructions configured for execution by the one or more hardware processors, the instructions being comprised in: an input module ( 120 a ) configured to: receive a plurality of images, at one or more time intervals, pertaining to a context under consideration, the received images being correlated and associated with at least one of a spatial and temporal information; and identify and transmit, at the one or more time intervals, at least a subset of the received images based on the spatial or temporal information and an adaptive threshold; a pre-processing module ( 120 b , 130 b ) configured to pre-process the received images, by performing at least one of (i) a first level of pre-processing, to enhance quality of the received images, by performing one or more of normalization, Principal Components Analysis (PCA) whitening, brightness correction, standardization and segmentation; and (ii) a second level of pre-processing, to adapt the received images for forward passing to the neural network model, by performing one or more of rotation, cropping, shifting, scaling and zooming; a feature extractor ( 120 c , 130 c ) configured to extract features from the received images, by forward passing the received images through a neural network model pre-trained on a dataset of a plurality of images pertaining to varied contexts; a clustering module ( 130 e ) configured to perform a first level of inferencing by clustering the received images into one or more classes using the extracted features, the step of clustering comprising: determining an optimal number of the one or more classes using a Silhouette coefficient; identifying the one or more classes based on similarity detected between the received images by performing at least one of: computing a first distance measure represented by a spherical distance of each of the received images with every other image in the received images; computing a likelihood of each image in the received images to belong to a class using a class distribution based on a Maximum A Posteriori probability (MAP); and computing a second distance measure represented by a spherical distance between each of the received images and a centroid of each of the one or more classes; wherein the number of the one or more classes equals the determined optimal number; and validating the quality of the one or more classes using one or more of an Normalized Mutual Information (NMI) score, a Rand Index and a purity measure; and associating the one or more classes with a tag based on the context under consideration.
This system processes a series of images captured over time to analyze and classify them based on spatial and temporal correlations. The system receives images associated with spatial or temporal metadata and selects a subset using an adaptive threshold. Pre-processing enhances image quality through normalization, PCA whitening, brightness correction, standardization, and segmentation, followed by adjustments like rotation, cropping, and scaling to prepare images for neural network analysis. A pre-trained neural network extracts features from the images, which are then clustered into classes using a multi-step process. The clustering determines the optimal number of classes via the Silhouette coefficient and groups images based on similarity, using spherical distance metrics, Maximum A Posteriori probability, and centroid-based distance calculations. Class quality is validated using Normalized Mutual Information, Rand Index, and purity measures, and the resulting classes are tagged according to the context. The system automates image classification for applications requiring temporal or spatial analysis, such as surveillance, medical imaging, or environmental monitoring.
10. The system of claim 9 , wherein the input module is further configured to obtain at least some metadata associated with the received images; and update the metadata associated with the received images based on the associated tags.
This invention relates to an image processing system designed to enhance image organization and retrieval by automatically tagging images and updating associated metadata. The system addresses the challenge of efficiently categorizing and retrieving images in large datasets, where manual tagging is time-consuming and inconsistent. The system includes an input module that receives images and obtains metadata associated with them, such as timestamps, locations, or user-generated tags. The input module then updates this metadata by adding or modifying tags based on the content of the images. For example, if an image contains a recognizable object or scene, the system can automatically assign relevant tags (e.g., "beach," "sunset") to improve searchability. Additionally, the system may include an analysis module that processes the images to extract features or patterns, which are then used to generate or refine tags. This ensures that the metadata remains accurate and up-to-date, even as new images are added or existing ones are modified. The system may also include a storage module to store the tagged images and their updated metadata, enabling efficient retrieval and organization. By automating the tagging and metadata updating process, the system reduces the need for manual intervention, improving efficiency and consistency in image management. This is particularly useful in applications like digital asset management, social media platforms, or surveillance systems where large volumes of images must be organized and searched quickly.
11. The system of claim 9 , wherein the feature extractor is further configured to append the features extracted by the neural network model, with additional features including morphological features and color related features to derive a master feature set; and compress the master feature set using dimensionality reduction methods.
The system relates to image processing and computer vision, specifically for enhancing feature extraction in image analysis. The problem addressed is the need for more comprehensive and efficient feature representation in image recognition tasks, where traditional methods may lack robustness due to limited feature diversity or high computational overhead. The system includes a neural network model that extracts initial features from input images. These features are then augmented with additional morphological features, such as shape, texture, and structural characteristics, as well as color-related features, such as hue, saturation, and brightness. This combined set forms a master feature set, which provides a richer and more discriminative representation of the image content. To optimize computational efficiency and reduce redundancy, the master feature set is compressed using dimensionality reduction techniques, such as Principal Component Analysis (PCA) or autoencoders. This compression retains the most salient features while minimizing data dimensionality, improving processing speed and storage efficiency. The system is designed to enhance the accuracy and reliability of downstream tasks like object detection, classification, or segmentation by leveraging a more comprehensive and optimized feature representation.
12. The system of claim 11 further comprising a classification module ( 130 f ) configured to perform a second level of inferencing by classifying the one or more classes into one or more sub-classes, the step of classifying comprising: obtaining a plurality of pre-trained tagger models ( 130 a ) associated with one or more sub-classes corresponding to one or more classes pertaining to the context under consideration, wherein the pre-trained tagger models are trained by the master feature set; classifying the one or more classes from the first level of inferencing into the one or more sub-classes based on the plurality of pre-trained tagger models; computing a confidence level for the classified one or more sub-classes; re-training the plurality of pre-trained tagger models with the images associated with the confidence level below a pre-defined threshold to obtain a plurality of classification models ( 120 d , 130 d ); and creating a knowledge ontology of the classifications models based on the one or more classes, the one or more sub-classes therein and the inter-relationships thereof.
This invention relates to an advanced classification system for hierarchical inferencing in image analysis. The system addresses the challenge of accurately categorizing images into multiple levels of detail, improving precision in automated image recognition tasks. The system includes a classification module that performs a second level of inferencing by further subdividing previously identified classes into sub-classes. This module utilizes a set of pre-trained tagger models, each associated with specific sub-classes relevant to the context being analyzed. These models are initially trained using a master feature set derived from the images. The classification process involves applying these tagger models to the classes identified in the first level of inferencing, assigning them to the appropriate sub-classes. The system then evaluates the confidence level of these classifications. For sub-classes with confidence levels below a predefined threshold, the corresponding tagger models are re-trained using the associated images to improve accuracy. The refined models are then integrated into a knowledge ontology, which organizes the relationships between the main classes, their sub-classes, and their interconnections. This hierarchical approach enhances the system's ability to provide detailed and contextually relevant image classifications.
13. The system of claim 12 , wherein the classification module is further configured to evaluate the received images associated with the confidence level below the pre-defined threshold with the associated metadata, prior to re-training the pre-trained tagger models.
This invention relates to an image classification system that improves accuracy by evaluating low-confidence classifications before retraining machine learning models. The system addresses the challenge of unreliable image tagging when automated classifiers produce uncertain results, which can degrade performance over time. The core system includes a pre-trained tagger model that processes images and assigns tags based on learned patterns. When the model generates a classification with a confidence level below a predefined threshold, the system triggers an evaluation process. This process examines the low-confidence images alongside their associated metadata, such as timestamps, source information, or user annotations, to determine whether the uncertainty stems from data quality issues, model limitations, or other factors. The insights from this evaluation guide the retraining of the tagger models, ensuring they improve accuracy without reinforcing incorrect classifications. The system dynamically adjusts its learning process by prioritizing problematic cases, reducing the risk of propagating errors. This approach enhances the reliability of automated image tagging in applications like content moderation, medical imaging, or surveillance, where accuracy is critical. The invention focuses on refining machine learning workflows by incorporating metadata-driven validation before model updates.
14. The system of claim 13 further comprising a database ( 130 g ) configured to store the received images and the metadata associated thereof; the one or more classes; the one or more sub-classes, the plurality of classification models and the knowledge ontology.
This invention relates to a system for image classification and knowledge management. The system addresses the challenge of efficiently organizing and retrieving images based on their content and associated metadata. The system includes a database that stores received images along with their metadata, classification models, and a knowledge ontology. The database also stores predefined classes and sub-classes used for categorizing the images. The classification models are trained to analyze the images and assign them to the appropriate classes and sub-classes. The knowledge ontology provides a structured framework for organizing and relating the classified images, enabling advanced querying and retrieval. The system enhances image management by integrating metadata, classification, and semantic relationships, improving searchability and usability in applications such as digital libraries, medical imaging, or surveillance systems. The database serves as a centralized repository, ensuring consistency and accessibility of the stored data and models. This approach streamlines image classification and retrieval, reducing manual effort and improving accuracy in large-scale image datasets.
15. The system of claim 14 , wherein the input module is further configured to identify and transmit at least a subset of the received images by performing one or more of: determining whether the received images are valid based on an entropy value associated thereof; and comparing the received images with a pre-determined number of previous images to determine whether the received images fit into the one or more sub-classes associated thereof, wherein the pre-determined number of previous images is the adaptive threshold based on the context under consideration.
This invention relates to an image processing system designed to filter and classify images based on entropy and contextual relevance. The system addresses the challenge of efficiently managing large volumes of images by dynamically validating and categorizing them to reduce processing load and improve accuracy. The system includes an input module that receives images and performs validation checks. It determines image validity by calculating an entropy value, which measures the randomness or information content of the image. Images with entropy values below a certain threshold may be deemed invalid or irrelevant. Additionally, the input module compares incoming images with a predefined number of previously processed images to assess whether they belong to specific sub-classes. The number of previous images used for comparison is adaptively adjusted based on the context, such as the environment or application in which the system operates. This adaptive threshold ensures that the classification remains relevant and efficient under varying conditions. The system also includes a processing module that further analyzes the validated images, extracting features and performing classification tasks. The adaptive threshold mechanism allows the system to dynamically adjust its filtering criteria, improving performance in real-time applications where image content and context may change frequently. This approach enhances computational efficiency and ensures that only relevant images are processed, reducing unnecessary workload and improving overall system accuracy.
16. The system of claim 15 , wherein the system is configured in a cloud-edge topology ( 100 ) having (i) the input module ( 120 a ) serving as an IoT gateway; (ii) the clustering module ( 130 e ), the classification module ( 130 f ), the pre-trained tagger models ( 130 a ) and the database ( 130 g ) are implemented as cloud ( 130 ) based devices; (iii) the pre-processing module ( 120 b ) is implemented both as a cloud end device as well as an edge ( 120 ) end device; and (iv) the feature extractor ( 120 c , 130 c ) and the classification models ( 120 d , 130 d ) are implemented as cloud end devices and are deployed on the edge end such that the edge end is updated with a current version thereof.
This invention relates to a cloud-edge computing system for processing and analyzing data, particularly in Internet of Things (IoT) environments. The system addresses the challenge of efficiently handling large-scale data processing by distributing tasks between cloud and edge devices to optimize performance, reduce latency, and improve scalability. The system includes an input module functioning as an IoT gateway to collect data from various sources. A pre-processing module processes this data, implemented both in the cloud and at the edge to enable local filtering and reduction before transmission. Feature extraction and classification models are deployed in the cloud but also updated on edge devices to ensure real-time processing capabilities. A clustering module and a classification module, along with pre-trained tagger models and a database, are hosted in the cloud for centralized analysis and storage. The system's cloud-edge topology ensures that computationally intensive tasks are handled in the cloud, while edge devices perform preliminary processing and classification, reducing bandwidth usage and latency. This architecture enhances efficiency, scalability, and responsiveness in IoT applications.
17. The system of claim 9 , wherein the clustering module is further configured to perform the first level of inferencing within one or more of the one or more classes.
A system for data analysis and classification processes data by organizing it into one or more classes. The system includes a clustering module that groups data points within these classes based on similarity or other criteria. The clustering module performs a first level of inferencing, which involves analyzing the grouped data to derive insights, patterns, or relationships. This inferencing step may include statistical analysis, pattern recognition, or other computational techniques to extract meaningful information from the clustered data. The system may also include additional modules for preprocessing data, such as normalization or filtering, to ensure the data is suitable for clustering. The clustering module operates within the predefined classes, refining the data organization to improve the accuracy and relevance of the derived insights. The system may be applied in various domains, such as machine learning, data mining, or business intelligence, where structured data analysis is required. The inferencing step enhances the system's ability to provide actionable insights from the clustered data, improving decision-making processes.
18. A computer program product comprising a non-transitory computer readable medium having a computer readable program embodied therein, wherein the computer readable program, when executed on a computing device, causes the computing device to: receive a plurality of images, at one or more time intervals, pertaining to a context under consideration, the received images being correlated and associated with at least one of a spatial and temporal information; identify and transmit, at the one or more time intervals, at least a subset of the received images based on the spatial or temporal information and an adaptive threshold; extract features from the received images, by forward passing the received images through a neural network model pre-trained on a dataset of a plurality of images pertaining to varied contexts, wherein the step of extracting features from the received images is preceded by a step of pre-processing comprising at least one of (i) a first level of pre-processing, to enhance quality of the received images, by performing one or more of normalization, Principal Components Analysis (PCA) whitening, brightness correction, standardization and segmentation; and (ii) a second level of pre-processing, to adapt the received images for forward passing to the neural network model, by performing one or more of rotation, cropping, shifting, scaling and zooming; perform a first level of inferencing, by clustering the received images into one or more classes using the extracted features, wherein clustering is performed by: determining an optimal number of the one or more classes using a Silhouette coefficient; identifying the one or more classes based on similarity detected between the received images by performing at least one of: computing a first distance measure represented by a spherical distance of each of the received images with every other image in the received images; computing a likelihood of each image in the received images to belong to a class using a class distribution based on a Maximum A Posteriori probability (MAP); and computing a second distance measure represented by a spherical distance between each of the received images and a centroid of each of the one or more classes; wherein the number of the one or more classes equals the determined optimal number; and validating the quality of the one or more classes using one or more of an Normalized Mutual Information (NMI) score, a Rand Index and a purity measure; and associating the one or more classes with a tag based on the context under consideration.
This invention relates to a computer program for analyzing image data to extract meaningful insights from varied contexts. The system receives multiple images at different time intervals, each associated with spatial or temporal metadata. It dynamically selects a subset of these images based on adaptive thresholds and spatial/temporal relevance. The images undergo a two-stage preprocessing pipeline: the first stage enhances image quality through normalization, PCA whitening, brightness correction, standardization, and segmentation, while the second stage adapts them for neural network processing via rotation, cropping, shifting, scaling, and zooming. A pre-trained neural network extracts features from these preprocessed images. The system then performs clustering to group images into classes, determining the optimal number of clusters using the Silhouette coefficient. Clustering relies on spherical distance calculations, Maximum A Posteriori probability-based class likelihoods, and centroid-based distance measures. The quality of these clusters is validated using Normalized Mutual Information, Rand Index, and purity metrics. Finally, the system tags the clusters based on the context of the analyzed images. This approach enables automated, context-aware image categorization and analysis.
Unknown
June 9, 2020
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.